This was a prospective, multicenter, proof-of-concept phase IIa clinical study in patients with severe infections conducted in collaboration between four study sites in Athens (Greece) and The BioArte Limited (San Gwann, Malta). A schematic representation of the study is represented in Figure 1 . The study was approved by the Ethics Committees of the participating study sites ( Supplementary Table 1 ). The study was conducted in accordance with the principles of the Declaration of Helsinki, Good Clinical Practice (GCP), and applicable regulatory requirements ClinicalTrials.gov ID NCT06141395 (sponsor Hellenic Institute for the Study of Sepsis).

Written informed consent was obtained from each patient or their legally designated representative prior to the initiation of any study procedures. Eligible participants included admitted adult patients of both sexes with a high suspicion of infection, defined by the presence of at least one sign of qSOFA (quick Sequential Organ Failure Assessment) score and for whom blood cultures were clinically indicated. Patients who had received any dose of antibiotics for the current infection episode, pregnant and breast-feeding women were excluded. During follow-up, data collected per patient included severity scores qSOFA and SOFA, Charlson’s comorbidity index, blood tests (total blood cell count, biochemistry and blood gases), infection type, and sepsis status (Sepsis-3 definition (Singer et al., 2016)). At the end of the study, SoC culture results were adjudicated by two blinded infectious disease experts using full clinical data to determine true pathogens versus putative contaminants.

For each enrolled patient, 20ml of whole blood were drawn for three consecutive days. The first sampling was performed before administration of the first dose of antibiotics. Blood samples were inoculated into one flask with growth medium for culture (Becton, Dickinson and Company, Sparks, USA) and immediately transported via courier within one hour to the central laboratory. Each flask was then placed into one autoincubator (BD) at 37°C. After six hours of incubation an aliquot of 5.5ml was removed under sterile conditions, stored at -80°C and shipped to The BioArte Ltd laboratories (San Gwann, Malta) for NGS analysis. Remaining blood culture (~14.5ml) was kept at 37°C and monitored for microbial growth to identify bacterial species and perform AST.

Growth flasks (BACT/ALERT^® FA Plus; bioMérieux, Marcy-l’Étoile, France) were incubated at 37°C in a BACT/ALERT^® VIRTUO^® automated incubator (bioMérieux, Marcy-l’Étoile, France). Time to positivity, defined as the interval between the initiation of incubation and detection of bacterial growth by the automated system, was recorded. Upon detection of growth, the flask content was subcultured on COLUMBIA Sheep Blood Agar plates (Bioprepare Microbiology, Attica, Greece) and used for microbial identification and AST using the BD Phoenix™ M50 Automated Microbiology System (Becton, Dickinson and Company, Sparks, USA) with the following components: BD Phoenix™ ID Broth, BD Phoenix™ PMIC/ID-88, and BD Phoenix™ NMIC/ID-503. If any signal of bacterial growth was recorded after 7 days, the flask content was aspirated and cultured to exclude the presence of slow-growing pathogens.

Total DNA was isolated from 0.5ml whole blood after 6 hours of incubation using the MagMax Microbiome Ultra II kit (Applied Biosystem, Waltham, Massachusetts, USA) and King Fisher device (Thermo Fisher Scientific). Samples were lysed through a bead-beating approach using the homogenizer MP FastPrep-24 5G (MP Biomedical) as follows: two cycles at 6.5 m/s for 45 sec each with a 5 minutes interval during which the samples were stored at 4°C. Two sequencing libraries based on Oxford Nanopore technology were prepared. The full-length 16S rRNA amplicon sequencing library was prepared as previously described (Pinzauti et al., 2024), introducing minor adjustments. Amplification was carried out in 25µL total volume pipetting 12.5µL Q5 Hot-Start High Fidelity 2x Master Mix (New England Biolabs), 4.5µL crude DNA extract, 400nM primers mix (Pinzauti et al., 2024), and 4µL nuclease-free water. The PCR reaction was carried out in a T100 thermal cycler (BioRad) using the following program: initial denaturation at 95°C for 3 minutes; followed by 32 cycles at 95°C for 20 s, 55°C for 30 s and 72°C for 2 minutes; and a final extension 72°C (5 minutes). PCR products were run on a 2% Agarose gel and cleaned with 0.6x Agencourt AMPure XP beads (Beckman Coulter). Barcoding PCR (sample multiplexing) was performed as follows: 12.5µL Q5 Hot-Start High Fidelity 2x Master Mix were mixed with 0.5µL phosphorylated barcoding primers mix (Pinzauti et al., 2024) and 100fmol 16S-amplicons from previous reaction (final reaction volume 25µL). The reaction was incubated at 95°C for 3 minutes (initial denaturation) followed by 15 cycles at 95°C for 15 s, 62°C for 15 s and 72°C for 2 minutes; and a final extension 72°C (5 minutes). Barcoded amplicons were purified with 0.6x AMPure XP beads, pooled together (equimolar pool), and adapter ligated following SQK-LSK114 (Oxford Nanopore Technologies) instructions. Briefly, 12.5µL Ligation Buffer (LNB), 5µL Quick T4 DNA Ligase (New England Biolabs), and 2.5µL Ligation Adapter (LA) were added to pooled barcodes (500ng DNA in 30µL). The reaction was incubated 10 minutes at room temperature. Clean-up was performed using 0.4x AMPure XP beads (AXP) washing twice with 250µL of Short Fragment Buffer (SFB) and resuspending in 15µL elution buffer (EB). Finally, 20fmol of DNA library were gently loaded in a dropwise manner onto a R10.4.1 flow cell (Oxford Nanopore Technologies) and sequenced using GridION Mk1b device (Oxford Nanopore Technologies). A minimum sequencing runtime of 2h was necessary to generate sufficient data.

The 16S-positive samples, identified by a sharp and clean band of ~1,500 bases in length during gel electrophoresis, were then subjected to Metagenomic sequencing. Metagenomics library was prepared following the Rapid PCR barcoding kit SQK-PRB114.24 (Oxford Nanopore Technologies). In short, 3µL crude DNA extract (expected 1 to 5ng DNA) were mixed with 1µL Fragmentation Mix (FRM) and incubated 2 minutes at 30°C followed by 2 minutes at 80°C (T100 thermal cycler, BioRad). Barcoding PCR reaction was prepared mixing 25µL LongAmp Taq 2x MasterMix (New England Biolabs), 20µL nuclease-free water, 4µL Tagmented DNA, and 1µL Rapid Barcode Primer mix (RLB, 10µM) in a clean 0.2ml thin-walled PCR tube. Reaction thermal profile consisted of an initial denaturation at 95°C for 3 minutes, followed by 25 cycles at 95°C for 15 s, 56°C for 15 s and 65°C for 6 minutes; and a final extension at 65°C for 6 minutes. Reaction was stopped by adding 4µL EDTA and incubating for 5 minutes (room temperature). Amplified samples were run on a 2% Agarose gel (expected sharp band of ~2,000 bases) and quantified using Qubit fluorometer. Barcoded samples were subsequently mixed generating an equimolar pool of 200–400 fmoles, purified with 0.6x AMPure XP beads (AXP), and resuspended in 15µL Elution Buffer (EB). Following, 10–50 fmol (20 to 100 ng DNA) of pooled library were transferred into a clean 1.5 ml Eppendorf DNA LoBind tube and mixed with 1µL diluted Rapid Adapter (RA), gently mixed and incubated for 5 minutes at room temperature. The library was gently loaded onto a R10.4.1 flow cell (Oxford Nanopore Technologies) mixing 37.5µL Sequencing Buffer (SB), 25.5µL Library Beads (LIB), and 12µL DNA library.

Raw sequencing data were live basecalled and demultiplexed using Guppy v. 6.1.5 and super accurate mode (quality threshold > 12) through the GridION Mk1b platform. Sequencing yields were evaluated using NanoPlot v. 1.41.0 (De Coster and Rademakers, 2023) measuring overall read distributions and statistics. 16S amplicons were analysed as previously described (Pinzauti et al., 2024). Briefly, sequencing reads were trimmed and filtered using cutadapt v. 3.5 (Martin, 2011) and dada2 v. 1.28.0 (Callahan et al., 2016), while chimeric reads were pruned using minimap2 v. 2.16 (Li, 2018) and yacrd v. 1.0.0 (Marijon et al., 2020) tools. Taxonomy was inferred employing emu v. 3.4.5 (Curry et al., 2022) and the Genome Taxonomy Database (GTDB, release v. 207, accessed on March 2023 (Parks et al., 2022)). Metagenomics data were first screened for human host DNA contamination. FastQ reads were mapped against reference human database T2T-CHM13 v2.0 (Nurk et al., 2022) with minimap2 (-ax map-ont preset, -f 1000) (Li, 2018); any mapped reads were removed using samtools v1.21 (samtools view -b -f 4) (Danecek et al., 2021). The remaining microbial reads were then taxonomically classified using Kraken2 v. 2.1.0 (Wood et al., 2019) and the PlusPF-16 database, a catalog of complete genomes from NCBI RefSeq including archea, bacteria, virus, fungi, fungi, protozoa and human DNA capped at 16GB (https://benlangmead.github.io/aws-indexes/k2, accessed on 9 October 2023). Classified species were further refined using Bracken v. 2.9 (Lu et al., 2017). Potential Antimicrobial Resistance (AMR) genes were predicted using abricate v. 1.0.1 (https://github.com/tseemann/abricate): reads were de novo assembled using Flye v. 2.9.5 (–meta preset) (Kolmogorov et al., 2020) and resulting contigs were mapped against The Comprehensive Antibiotic Resistance Database (Jia et al., 2017). Detected AMR genes were further screened for clinical relevance: only genes conferring resistance to antimicrobials used in clinical and hospital settings, as defined by the World Health Organization Medically Important Antimicrobials List for Human Medicine (WHO MIA, https://cdn.who.int/media/docs/default-source/gcp/who-mia-list-2024-lv.pdf?sfvrsn=3320dd3d_2), were reported.

The study was not powered for the primary endpoint since this was a proof-of-concept phase IIa study. Since the rate of positivity of SoC cultures was expected to be 20% (Hadziyannis et al., 2004; Previsdomini et al., 2012), 100 patients were enrolled. Statistical analyses were performed with IBM SPSS Statistics for Windows (Version 29.0.2.0 Armonk, NY: IBM Corp) evaluating the classification matching rate between the 16S rRNA gene sequencing and the SoC cultures for patients classified with sepsis. Classification performance was evaluated by measuring the Sensitivity, the Specificity, the Positive Predictive Value (PPV), the Negative Predictive Value (NPV) and the Accuracy of the 16S rRNA sequencing approach. The aforementioned values were calculated as the ratios of True Positives/(True Positives + False Negatives), True Negatives/(True Negatives +False Positives), True Positives/(True Positives + False Positives), True Negatives/(True Negatives + False Negatives) and (True Positives + True Negatives)/(True Positives + True Negatives + False Positives + False Negatives), respectively. For the evaluation of diagnostic performance, results from SoC cultures were used as the reference standard. SoC culture results with disagreements between adjudicators were excluded from analysis. The above-mentioned values were expressed with percentages, rounded to the first digit after decimal point, accompanied by their 95% Confidence Intervals (CIs). For the secondary endpoints, the time comparison was assessed with the Mann-Whitney U test, with the significance level at 5%. Continuous variables following normal distribution were expressed by the mean and SD. Continuous variables not following the normal distribution were expressed by the median and interquartile range (Q1-Q3). Binomial variables were expressed as absolute and percentage frequencies with 95% CIs.

Patients (n=262) with suspected sepsis were assessed for eligibility from November 21st, 2023 to March 15th, 2024. Of these, 162 patients were excluded as they received antibiotic treatment. The remaining 100 patients were enrolled in this study, including 40 males and 60 females, with a median (Q1-Q3) age of 79 (68–86) years. The medical history conditions (comorbidities) reported for the study population were diabetes mellitus type 1 or 2 (29 patients), heart failure (16 patients), chronic renal disease (13 patients), chronic obstructive pulmonary disease (19 patients), dementia (28 patients), coronary heart disease (20 patients), atrial fibrillation (25 patients), depression or psychosis (20 patients), residency in long-term care facility (14 patients), and intake of antimicrobials for any reason the last three months (19 patients). The median (Q1-Q3) Charlson’s comorbidity index score registered for all patients was 5. Regarding the type of infection, 8 patients were enrolled with primary bacteremia, 56 patients with community-acquired pneumonia (CAP), 2 patients with hospital-acquired pneumonia (HAP), 9 patients with acute biliary tract infections and 25 patients with acute pyelonephritis. The median (Q1-Q3) SOFA score at day 1 was 3 (Kumar et al., 2006; Giamarellos-Bourboulis et al., 2017; Lamy et al., 2020; Rudd et al., 2020; Shi et al., 2024). In total, 71 out of 100 enrolled patients met Sepsis-3 definition.

SoC blood cultures detected 32 positives over the three consecutive follow-up days. According to both adjudicators, 11 of these cases involved skin commensals rather than true pathogens and were excluded. Additionally, the adjudicators disagreed on 2 cases and therefore were discarded for the primary endpoint analysis. A further 5 samples were excluded as the patients did not meet the Sepsis-3 criteria for sepsis positivity. In total, true pathogens were detected in 14 patients by SoC cultures ( Figure 2 ). The list of positive samples is provided in Table 1 . According to microbiological data, a panel of 17 antimicrobials was tested evaluating sensitivity or resistance profiles. Because patient BAB002 failed species-level classification, it was not processed for AST. The list of tested antimicrobial compounds and their relative resistance profiles are listed in Supplementary Table 2 .

PISTE full-length 16S rRNA sequencing successfully detected 13 out of 14 SoC positives ( Table 1 ; Figure 3 ). Patient BAB002 failed proper 16S gene amplification resulting in the absence of a visible, clear fluorescent band during gel electrophoresis. Nevertheless, the sample was processed through the entire sequencing library preparation as described but no sequencing data were generated, thus failing pathogen identification. Interestingly, PISTE workflow detected the presence of two microbial pathogens in patients BAB037 and BAB042. Specifically, in patient BAB037 the NGS-based analysis identified the pathogens Klebsiella pneumoniae (concordant with SoC cultures) and Acinetobacter baumannii (undetected by SoC cultures), while Providencia stuartii (concordant) and Staphylococcus hominis (undetected) were identified in patient BAB042. The overall concordance between the PISTE and SoC cultures demonstrated an overall 95.7% Accuracy (95% CI: 88-98.5), 91.7% Sensitivity (95% CI: 64.6-98.5), 96.5% Specificity (95% CI: 88.1-99), 84.6% PPV (95% CI: 57.8-95.7), and 98.2% NPV (95% CI: 90.6-99.7).

The 16S-positive samples were then processed through Metagenomics strategy prior antimicrobial resistance prediction. Of the 13 positive samples, only 7 (BAA002, BAB029, BAB037, BAB041, BAB042, BAB046, and BAD003) were successfully amplified and generated sufficient sequencing data. The remaining samples (6/13) likely failed due to insufficient microbial DNA, reflecting a low microbial biomass below technology detection threshold. Bracken classification showed 100% agreement with 16S-based taxonomy, further highlighting the presence of two microbial species in samples BAB037 and BAB042. Detected AMR genes (CARD database) were manually screened for the presence of relevant genes and compared to standard AST results for the tested antimicrobial compounds including Aminoglycosides, Cephalosporins, Carbapenems, Penicillins, Quinolones, and Tetracyclines ( Figure 4 ). Resistance to Penicillins and Cephalosporins was marked by the detection of any of the following bla _TEM-1, bla _CTX-M-15, bla _CTX-M-27, bla _CTX-M-139, bla _OXA-1, bla _OXA-10, bla _OXA-66, bla _SHV-182, or bla _Z genes. Carbapenem resistance was predicted based on the identification of Metallo-β-lactamases bla _NMD-1 and bla _VIM-1, or other carbapenemases like bla _KPC and bla _OXA, or multidrug efflux pump mexA and mexB genes. Detection of genes ACT-like aac(6’)-lb10, aac(6’)-lb-cr, and NUT-like aadA2, aadA11 were used to predict Aminoglycoside resistance. Tetracycline resistance was inferred by the identification of tetK gene and efflux pump mexA and mexB genes, while Quinolone resistance was identified through detection of qnrA, qnrB, and qnrS genes. Respective cross-tabulations of the two assays for AST are presented in Tables 2A-E ( Figure 4A ). Antimicrobial resistance prediction showed 100% Accuracy for Cephalosporins and Penicillins (95% CI: 64.6-100), 85.7% for Carbapenems (95% CI: 48.7-97.4), and 71.4% for Aminoglycosides (95% CI: 35.9-91.8). Lower precision was obtained for Tetracyclines and Quinolones resistances (57.1% Accuracy, 95% CI: 25-84.2) ( Figure 4B ). The analysis of the turnaround times revealed a significant benefit in favor of the PISTE technology (P value < 0.0001). The NGS-based assay provided pathogen identification and AST at a fixed timepoint of 12.0 hours, while SoC cultures required a median of 30.4 hours (Q1-Q3: 28.9-33.7) to pathogen identification and AST reporting. However, reported time ranges were approximate as actual turnaround times are influenced by sequencing runtime, microbial load, and available computational resources (data interpretation).